In Two Minds about Lifelong Learning: Exploring Hemispheric Redundancy and Specialisation in Neural Models
Authors: Benjamin Smith, Levin Kuhlmann, Kaushik Roy, Gideon Kowadlo
Organizations: Data Science and AI, Monash University, Wellington Road, Clayton, 3800, Victoria, Australia · CSIRO Robotics, CSIRO, Australia · Cerenaut, Victoria, Australia
Persistent intelligent systems require the ability to learn continually, but current machine learning approaches face significant challenges in this area compared to biological learning systems. Machine learning algorithms typically trade off retention of previously learned information and adaptation to new or changing data patterns. When continual learning capabilities are absent, algorithms must undergo retraining using the entire data set, an approach that becomes impractical when original training data are unavailable due to storage constraints, financial or computational costs, or privacy restrictions. However, biological animals can learn continually, without experiencing catastrophic forgetting. This paper attempts to build a high-level framework for how animals learn and preserve knowledge by modelling neural components and states that are known to be related to memory consolidation. We focus on three concepts: experience replay, REM sleep, and bilaterality. We propose 4MAS (4 Module Awake/Sleep), a novel macroarchitecture demonstrating how machine learning models might benefit from asymmetric hemispheres, each with their own long- and short-term memory mechanisms, and how a period of sleep between incremental learning tasks might benefit memory consolidation. Finally, we present results showing that our architecture achieves statistical improvements over established generative replay baselines on Split-MNIST (98.3%) and Split-Fashion-MNIST (84.9%), while reducing forgetting by more than half relative to the strongest replay baseline and demonstrating monotonic capacity scaling on Split-CIFAR-100 (29.29%).
Figures & tables
Figure 1: LTM inputs and outputs: (a) LTM training on image-label pairs for each hemisphere during awake and sleep phases. (b) Sample generation with temperature τ from each hemisphere provides two synthetic input sources for subsequent LTM training tasks. (c) During inference, both hemispheres process inputs and the highest confidence classification is selected.
Figure 2: MNIST datasets are split into 5 tasks of 2 classes each for the Class-IL scenario: (a) Split-MNIST, (b) Split-Fashion-MNIST.
Figure 3: Awake/sleep training phases: (a) Awake training: LTMs learn new task data interleaved with STM samples. (b) Sleep training: LTMs are fine-tuned on contralateral representations.
Figure 4: 2-dimensional mapping of the LTM’s latent space after training. τ scales the sampling radius in units of the prior standard deviation. High temperature τ>1 increases the diversity of output representations (16 latent dimensions used in final configurations).
Dataset
Method
Final Acc
Forg.
BWT
FWT probe ( Δ )
Drift
(%) ↑
(%) ↓
(%) ↑
( % ) ↑
↓
Split-MNIST
Joint (145M)
98.14±0.12
—
—
—
—
4MAS (Our method, 145M)
98.30±0.04
1.2
−1.2
91.5(−5.7)
1.0
B-IR (5M)
93.5±0.2
6.4±0.2
−6.4±0.2
98.7(+1.5)
0.2±0.0
GR (186M)
94.50±0.49
8.9±0.4
−8.9±0.4
97.6(+0.5)
42.1±1.8
LwF (1.9M)
24.2±0.4
0.0
0.1
96.7(−0.5)
15.2
Table 1: Class-IL evaluation across Split-MNIST, Split-Fashion-MNIST, and Split-CIFAR-100 benchmarks. For each method, we report performance across capacity sweeps, with total trainable parameter counts indicated in parentheses. Best values (excluding the Joint upper bound) are in bold. Metrics display mean and standard error of the mean ( ±SEM ) across random seeds after learning all tasks. FWTprobe denotes representation linear probe accuracy on future tasks and relative transfer ( Δ ) above random initialisation.
Figure 5: (a) Performance variation with sleep phase learning rates (LR) as a percentage of awake phase LR, Split-Fashion-MNIST. (b) Dominance scores categorised as RH (<50%), symmetrical (50%), or LH (>50%). Memory consolidation during sleep phase increased LH dominance, while new task learning during awake phase increased RH dominance; produced by lateralisation techniques (class representation bias by age in LH and asymmetric generator temperatures).
Figure 6: Fashion-MNIST: (a) Hemispheric accuracy for each task. (b) τ -asymmetry results, with and without sleep training phase.
τ - right
1
2
3
4
5
τ - left
1
0.6301
0.7122
0.7445
0.7132
0.6984
2
0.7021
0.6895
0.6662
0.6342
0.6178
3
0.6768
0.6116
0.5907
0.5595
0.5504
4
0.6593
0.6128
0.5561
0.5551
0.5233
5
0.6453
0.576
0.5417
0.555
0.5148
Table 2: LH and RH generator temperature vs Split-Fashion-MNIST accuracy. Degraded performance observed for symmetrical temperatures.
CCS - RH
90%
70%
50%
30%
10%
0%
LSCC
CCS - LH
90%
0.7562
0.7662
0.7668
0.7668
0.7536
0.7505
0.7656
70%
0.7618
0.7724
0.7753
0.7725
0.7607
0.7589
0.7796
50%
0.7695
0.7706
0.7733
0.7667
0.7648
0.7573
0.7779
30%
0.7703
0.7718
0.7758
0.7627
0.7565
0.7247
0.7799
10%
0.7673
0.7704
0.7675
0.7509
0.7351
0.727
0.7757
Table 3: Comparison of mechanisms for selecting Fashion-MNIST memories for STM buffer storage. Storing latent space centroids via K-Means clustering (LSCC) produced the most consistent results on Split-Fashion-MNIST, whereas highly asymmetric confidence selection ( L=0.1/R=0.5 ) (CCS) achieved peak performance on the more complex Split-CIFAR-100 benchmark.
Figure 7: (a) Biasing class representation by recency; weighting training heavily toward older classes significantly boosts early-class accuracy with only minor performance degradation on recent classes. (b) Effect of STM memory buffer size on accuracy; gains from larger buffers plateau after 200 samples, while also diminishing the relative impact of the sleep phase.
Figure 8: Capacity scaling curves comparing our bilateral 4MAS architecture against the unilateral B-IR baseline across (a) Split-MNIST, (b) Split-Fashion-MNIST, and (c) Split-CIFAR-100. The plots show final accuracy as a function of total model parameters (in millions).
Figure 9: Ablation performance for Split-MNIST and Split-Fashion-MNIST. Configurations including the sleep phase consistently perform best on both datasets, where sleep informed hemispheric confidence levels lead to system accuracy exceeding individual hemisphere accuracies.
Class-IL, buffer 5000, SWS/REM epochs 75/50, LR 0.0022
fourmas_cifar100_multi_arch
4
3.77
Pre-trained tap layer 2 feature extractor, latent dim 64
fourmas_cifar100_scratch_sota
15
6.85
Scratch training (no ImageNet), buffer 10000, latent dim 128
Appendix
Table 5: Best configurations and parameters across Split-CIFAR-100 Optuna sweeps.
ID
Model Type
SupCon Weight
LH Freeze
BSS Active
Class-IL (%)
Task-IL (%)
1
vae
0.0
False
False
1.52
12.04
2
vae
0.0
False
True
1.02
9.61
3
vae
0.0
True
False
1.19
11.17
4
vae
0.0
True
True
1.13
9.26
5
vae
0.5
False
False
2.13
17.33
6
vae
0.5
False
True
2.01
14.90
Appendix
Table 6: Grid search performance over generative model types, supervised contrastive (SupCon) weights, left-hemisphere parameter freezing, and BSS active gating.
Selection Strategy (LH / RH)
Left Centroid ( L )
Right Centroid ( R )
Pre-training Tasks
Synaptic Intelligence (SI)
Class-IL Acc (%)
CCS (Symmetric Baseline)
0.5
0.5
1
No
21.38
CCS (Symmetric)
0.5
0.5
2
No
25.34
CCS (Symmetric)
0.5
0.5
3
No
29.18
CCS (Symmetric)
0.5
0.5
4
No
30.46
CCS (Symmetric)
0.5
0.5
5
No
29.01
CCS (Left-Skewed)
0.7
0.3
1
No
21.78
Appendix
Table 7: Split-CIFAR-100 performance across selection strategies (CCS, LSCC, Hybrid), pre-training task horizons, and local SI regularisation.
Model Size
Architecture
Total Parameters
Class-IL Acc (%)
Forgetting (%)
14M Scale
4MAS (Ours, Scaled-down)
14.1M
13.16
25.8
B-IR (Standard Baseline)*
13.3M
21.0
61.1
35M Scale
4MAS (Ours, Scaled-down)
34.5M
18.41
25.8
B-IR (Ours, Scaled-down)
35.8M
23.01
58.9
70M Scale
4MAS (Ours, Champion)
70.0M
24.45±0.32
25.8±1.2
B-IR (Ours, Scaled)
62.7M
26.27±1.37
52.82±1.97
Appendix
Table 8: Split-CIFAR-100 Class-IL performance across parameter capacity scales.
Figure 10: Continual learning post-REM accuracy trajectories across class scale (10 to 100 classes) for top VAE configurations and Quality-Gated replay regimes, showing mean performance and standard deviation error bars across random seeds.
One of the critical limitations of artificial neural networks is their lack of ability to continually learn: training on new tasks often leads to interference and forgetting of the previous ones. While several algorithms have been proposed to protect old memories from interference, they are typically applied during or immediately after each new episode of training. In contrast, humans and animals can learn continuously, acquiring multiple new memories during active learning before consolidating all of them into long-term storage. Here we show that multiple new tasks can be trained sequentially before an unsupervised sleep-like replay phase is applied to partially restore performance across all previously learned tasks. Our study further suggests that task-specific information remains resilient to new training but decays gradually as network is trained on new tasks. These findings point to novel principles for developing a broad range of continual learning AI solutions.
Anthony Bazhenov, Jean Erik Delanois, Giri P. Krishnan
Khoury College of Computer Sciences, Northeastern University, Boston, MA · Department of Medicine, University of California, San Diego, La Jolla, CA, USA · ARTISAN, Georgia Institute of Technology, Atlanta, GA USA
Artificial Intelligence (AI) systems often perform well on isolated tasks but struggle under continual learning conditions, where training on new tasks can overwrite previously acquired knowledge, a failure mode known as catastrophic forgetting. Biological learning systems reduce this interference through complementary memory processes involving rapid hippocampal encoding and slower cortical consolidation. This study introduces NeuroSynth, a brain-inspired continual reinforcement learning architecture designed to mitigate catastrophic forgetting through a dual-pathway consolidation mechanism. NeuroSynth separates rapid task acquisition from long-term retention using distinct "plan" and "habit" pathways combined with replay and knowledge distillation. NeuroSynth was evaluated against Proximal Policy Optimization (PPO) and Elastic Weight Consolidation (EWC) across three sequential navigation tasks with changing goal locations in a non-revisitation continual learning setting. Across six independent seeds, NeuroSynth preserved substantially more early-task knowledge than PPO after sequential training, achieving 18.00% Task A success rate compared to 0.33% for PPO (p = 0.014929, Cohen's d = 1.49) and 35.33% Task B success rate compared to 0.00% for PPO (p = 0.002376, Cohen's d = 2.31). NeuroSynth also demonstrated higher final Task C performance than EWC, achieving 9.00% compared to 2.00% (p = 0.226643, Cohen's d = 0.56), indicating a moderate but not statistically significant advantage. These findings suggest that biologically inspired consolidation mechanisms may improve the stability-plasticity balance in continual reinforcement learning systems.
The past few decades have witnessed significant advances in the design of machine learning algorithms, from early studies on task-specific shallow models to more general deep Large Language Models (LLMs). Despite showing promising results in tasks that require instant prediction or in-context learning, existing models lack the ability to continually learn and effectively transfer their temporal in-context knowledge to their long-term parameters. Inspired by human learning process, we introduce a ''Sleep'' paradigm that allows the models to continually learn, distill their short-term fragile memories into stable long-term knowledge with replay, and recursively improve themselves with ''Dreaming'' process. In more detail, sleep consists of two stages: (1) Memory Consolidation: an upward distillation process, called Knowledge Seeding, where the memories of a smaller-self are distilled into a larger network to provide more capacity while preserving the knowledge. As a proof of concept, we present a new Generalized Distillation process for {Knowledge Seeding} (i.e., the combination of on-policy distillation with Reinforcement Learning (RL)-based imitation learning); (2) Dreaming: a self-improvement phase, where the model uses RL to generate a curriculum of synthetic data to rehearse new knowledge and refine existing capabilities without human supervision. Our experiments on long-horizon, continual learning, knowledge incorporation, and few-shot generalization tasks support the importance of the sleep stage.